Status: prototyped and measured, deliberately NOT landed. On hold by decision 2026-08-21.
The closed-form azimuth solve for snyder_inv_xyz (branch claude/snyder-closed-form, commit e0ac311) has an optional follow-up: replace the input decode's atan2 + sincos(Azps) with half-plane sign tests plus an exact ROT_SECTOR rotation of w. It was fully prototyped and benchmarked, and held for the reasons below — recorded here so the decision can be revisited with evidence instead of redone.
Measured (real level-12 corpus, 120,060 (f,w) pairs)
| kernel |
ns/call |
| Newton (superseded) |
220.9 |
| closed form (shipped on the branch) |
120.5 |
+ trig-free decode, sqrt(x²+y²) for ρ |
84.6 |
+ trig-free decode, hypot for ρ |
94.9 |
Why it's held
- Exceeds the approved output-change envelope. Cumulative deviation from the Newton-era kernel is 1.255e-14 on real data, but 1.722e-14 on a full-circle sweep including fringe radii — past the ≤1.62e-14 envelope approved for the closed-form change. Adopting it is a new accuracy decision (propose re-approving the fringe envelope at 2e-14).
- Sector-ray sensitivity is structural. IGeo7 corner offsets are exact Eisenstein arithmetic, so real lattice points land exactly on ±120° sector rays;
atan2 and the sign tests round a band of near-ray inputs into different sectors (20 mismatches in the 120,060-pair real corpus; 81/486 on exact rays). Measured to be benign — the sector fold is self-consistent, max ‖dxyz‖ across mismatching real points 9.37e-16 — but this equivalence should be pinned by a dedicated test if the decode lands, not just measured once.
- Underflow trap in the fast variant. The
sqrt(x²+y²) form underflows for |w| < ~1.5e-162 → division by zero → NaN vector. Unreachable from current lattice scales, but a trap. The hypot form is immune and costs 10.2 ns of the 35.9 ns win.
If picked up later
Take the hypot variant (94.9 ns, 1.27× on top of the closed form), re-approve the fringe deviation envelope to 2e-14, and add a test pinning the on-ray fold equivalence.
Full measurements and the prototype code: regrid-notes/2026-08-21-snyder-impl.md §8 (workspace notes) and the scratch harness referenced there.
🤖 Generated with Claude Code
https://claude.ai/code/session_01AC7rbi9eqZypWseMjBGSJ2
Status: prototyped and measured, deliberately NOT landed. On hold by decision 2026-08-21.
The closed-form azimuth solve for
snyder_inv_xyz(branchclaude/snyder-closed-form, commit e0ac311) has an optional follow-up: replace the input decode'satan2+sincos(Azps)with half-plane sign tests plus an exactROT_SECTORrotation ofw. It was fully prototyped and benchmarked, and held for the reasons below — recorded here so the decision can be revisited with evidence instead of redone.Measured (real level-12 corpus, 120,060
(f,w)pairs)sqrt(x²+y²)for ρhypotfor ρWhy it's held
atan2and the sign tests round a band of near-ray inputs into different sectors (20 mismatches in the 120,060-pair real corpus; 81/486 on exact rays). Measured to be benign — the sector fold is self-consistent, max ‖dxyz‖ across mismatching real points 9.37e-16 — but this equivalence should be pinned by a dedicated test if the decode lands, not just measured once.sqrt(x²+y²)form underflows for|w| < ~1.5e-162→ division by zero → NaN vector. Unreachable from current lattice scales, but a trap. Thehypotform is immune and costs 10.2 ns of the 35.9 ns win.If picked up later
Take the hypot variant (94.9 ns, 1.27× on top of the closed form), re-approve the fringe deviation envelope to 2e-14, and add a test pinning the on-ray fold equivalence.
Full measurements and the prototype code:
regrid-notes/2026-08-21-snyder-impl.md§8 (workspace notes) and the scratch harness referenced there.🤖 Generated with Claude Code
https://claude.ai/code/session_01AC7rbi9eqZypWseMjBGSJ2